Fixed-Line and Mobile Internet: Complements or Substitutes?

Transcription

1 Fixed-Line and Internet: Complements or Substitutes? Jiao (Tina) Xu PhD Candidate of Information Technology Management Georgia Institute of Technology Chris Forman Professor of Information Technology Management Georgia Institute of Technology Yu (Jeffrey) Hu Associate Professor of Information Technology Management Georgia Institute of Technology October 14 th, 2014 Abstract This article examines whether the quality of local fixed-line Internet service influences mobile Internet adoption and usage. An empirical analysis shows that local fixed-line Internet speed relates negatively to mobile Internet adoption and usage; if the local fixed-line connection is insufficient, consumers tend to get online through their mobile phones. Further, better local mobile Internet coverage increases the likelihood of adopting and using the mobile Internet. Neither fixed-line nor mobile Internet speed has significant impacts on mobile-specific offline services such as taking photos or videos. In some circumstances, substitution between the two platforms is stronger, such as among younger consumers and those living in areas with lower fixed-line Internet speeds. The authors discuss the implications of these findings for policy makers, consumers, and mobile service providers. Keywords: Internet, Fixed-Line Internet, Substitution, User Behavior, Econometrics

2 1 1. Introduction According to estimates by the International Telecommunication Union (2013), the 2.1 billion active mobile Internet users in the world at the end of 2013 represented 29.5% of the global population. Internet subscriptions grew 40% annually between 2010 and 2013, roughly three times the growth rate for fixed broadband subscriptions (International Telecommunication Union 2013). This rapid growth has been accompanied by increasing research interest in how consumers use the mobile Internet (e.g., Bang et al. 2013; Ghose and Han 2011; Ghose, Han and Xu 2013; Xu et al. 2014a; Xu et al. 2014b), though we still know relatively little about the interrelationships of different Internet platforms and their effects on consumption, including whether complementarity or substitution is more prevalent. On the one hand, fixed-line and mobile networks provide many of the same basic functionalities, so consumers may use the two platforms interchangeably. On the other hand, important differences demarcate the two Internet platforms, such that mobile service is distinguished by its mobility, whereas fixed-line service boasts superior service quality and reliability. Ghose, Goldfarb, and Han (2012) show that that consumers behave differently when using mobile versus fixed-line Internet connections. If people leverage the different advantages of each platform simultaneously, the value of consuming both platforms may be greater than the value of consuming each of them individually. Therefore, the question of whether fixedline and mobile Internet platforms are substitutes or complements requires empirical analysis. In turn, the substitutability or complementarity of fixed-line and mobile services has significant policy implications for the market structure of mobile service provision. Telecommunication policy makers worry about a lack of competition in the mobile service market, where service providers often operate with few competitors. The U.S. Department of Justice thus sought to block a proposed purchase of T- USA by AT&T, ultimately leading AT&T to abandon its bid (AT&T, 2011). However, if mobile networks substitute for fixed-line networks, the market structure actually alleviates many competitive concerns, because it suggests that both wired and wireless service providers compete with mobile service providers.

3 2 We offer a framework and analyze relevant data to understand how the quality of an incumbent platform (fixed-line Internet) influences the adoption and use of the emerging platform (mobile Internet). If the two platforms are substitutes, increases in the use of one platform will decrease the utility of adopting and using the other. If the two platforms are complements, increases in the use of one platform also increase the utility of adopting and using the other. Because an increase in the quality of one platform increases the utility of adopting and using that platform, we study how increases in the service quality of the fixed-line Internet affect the adoption and usage of the mobile Internet. For this investigation, we employ a novel survey data set of 28,117 U.S.-based mobile phone users in the second quarter of 2011, combined with local Internet coverage data from the National Broadband Map (NBM). We estimate a two-stage logit model with our cross-sectional data, such that consumers first choose whether to adopt a mobile data plan, and then choose whether to conduct online activities on their mobile phones. Our primary interest is in determining how the speed offered by local fixed-line and mobile service providers influences mobile Internet adoption and usage, so we adopt three approaches to assess whether the relationships are causal. First, we include area demographic variables as controls, because a plausible alternative explanation of the causal effect of Internet speed on mobile Internet adoption is that higher-density, higher-income, or higher-education areas encourage both Internet service quality and Internet adoption (e.g., Greenstein and Mazzeo 2006; Kolko 2012; Prieger 2003). Second, to investigate whether Internet speed influences mobile Internet adoption and usage through other omitted variables, we employ a control function approach (Imbens and Wooldridge 2007; Wooldridge 2014), with two instruments: the proxy costs of providing local wire service and historical mobile coverage information in adjacent areas. Third, we conduct a falsification analysis to test whether our effects appear when they should not. Although better local mobile Internet coverage can increase the likelihood of adopting and using mobile Internet plans, the impact of local fixed-line Internet speed on mobile Internet adoption and usage is negative. Thus, fixed-line and mobile Internet appear to be substitutes. However, we do not find

4 3 significant impacts of fixed-line or mobile Internet speed on the consumption of mobile-specific activities that do not require an Internet connection. The average effect also masks significant heterogeneity in consumer responses. In particular, younger consumers view the fixed-line and mobile Internet as closer substitutes. Because fixed-line and mobile Internet speeds are more comparable in areas with relatively low fixed-line Internet speeds, consumers living in such areas also exhibit stronger substitution behaviors. Overall, this study explicates cross-platform Internet consumption behavior on mobile devices. Such understanding is increasingly important as online search, browsing, and purchase behavior move more to mobile devices (comscore 2012), and it offers several substantive contributions to extant literature. Building on recent research on the interrelationships of different channels (e.g., Avery et al. 2012; Bang et al. 2013; Gentzkow 2007; Forman, Ghose, and Goldfarb 2009; Xu et al. 2014b), we seek to define when complementarity and substitution are most likely to occur across different Internet platforms. Although several recent works consider consumer behaviors on the mobile Internet (e.g., Ghose and Han 2011; Ghose, Goldfarb, and Han 2012), they focus mainly on consumers usage, after they already have adopted the mobile Internet. We seek to address a more fundamental question about how changes in fixed-line Internet service quality influence consumers decisions to adopt and then use the mobile Internet. Finally, by demonstrating substitution effects between the fixed-line and mobile Internet and identifying consumer segments for which substitution is greater, we provide insight for policy makers and practitioners about competition among Internet service providers, potential infrastructure investment strategies, and appropriate targeting. 2. Theoretical Framework Different factors influence the costs and benefits of adopting any general purpose technology (Bresnahan and Trajtenberg 1995). According to a standard probit model of diffusion (e.g., David 1969; Karshenas and Stoneman 1993), consumers adopt a technology if the benefits of doing so exceed its adoption costs. As the returns (i.e., benefits minus costs) on adopting increase, the likelihood of adoption increases too. Post-adoption usage then determines how much value consumers derive from the

5 4 technology (Fichman and Kemerer 1997). Similar to the adoption decision, consumers use the technology more if the returns from usage are higher, conditional on adoption. We investigate whether the fixed-line and mobile Internet are complements or substitutes by examining how fixed-line Internet quality affects the probability of mobile Internet adoption and usage of the mobile Internet. Following prior economics literature, we consider the two platforms complements when the utility of consuming them together is greater than the combined utilities of consuming them separately; they are substitutes if the utility of consuming the two platforms together is less than the combined utility of consuming them separately (Gentzkow 2007). Improvements in the quality of one platform increase the benefits, and hence increase the utility of adopting and using that platform, as long as costs remain the same. If the two platforms are substitutes, the utility of adopting and using one platform declines with the increasing quality of the other platform. In contrast, if two platforms are complements, the utility of adopting and using one platform increases with the quality of the other platform. Thus, if a fixed-line Internet with lower quality is associated with greater mobile Internet adoption and usage, the two platforms likely are substitutes. However, if the quality of local fixed-line Internet is positively associated with mobile Internet adoption and usage, it suggests complementarity between the two Internet platforms Probability of Adoption For both fixed-line and mobile Internet services, Internet connection speed is one of the most important determinants of service quality (Limayem, Khalifa, and Frini 2000; Yoo and Donthu 2001). We first discuss how mobile Internet speed affects the returns to mobile Internet adoption, and then focus on how fixed-line Internet speed affects the returns to mobile Internet adoption. Consumers obtain more benefits if they adopt a faster Internet service, so mobile Internet connection speeds should have a positive impact on the returns of adopting the mobile Internet service. Internet connection speed should be positively associated with the likelihood of mobile Internet adoption.

6 5 However, the question of whether the fixed-line and mobile Internet are complements or substitutes is an empirical one, because past evidence supports both relationships. Consumers may adopt an emerging platform to substitute for an existing platform if the emerging platform closely duplicates the capabilities of existing ones (Gentzkow 2007) or offers new capabilities (Alba et al. 1997). If consumers perceive the mobile Internet as similar to the fixed-line Internet along key dimensions, the returns from adopting the mobile Internet increase when local fixed-line Internet speed declines. In this case, the fixedline Internet connection speed would be negatively associated with the likelihood of mobile Internet adoption. However, consumers also might adopt an emerging platform to complement an existing platform if the platforms offer different comparative advantages (Avery et al. 2012; Peukert and Kretschmer 2014). For example, consumers can access the Internet through their mobile devices anytime and anywhere, whereas a fixed-line network imposes stricter limitations on geographical mobility and access. devices typically have much smaller screens than personal computers though, which may influence user behavior (Ghose, Goldfarb, and Han 2012). Noting these separate benefits, consumers may take advantages of both platforms simultaneously. For example, consumers can check instant news and information through the mobile Internet but use personal computers and a fixed-line network to search for in-depth information. Such complementarity is manifest in an empirical prediction: Fixed-line Internet connection speed is positively associated with the likelihood of mobile Internet adoption. These substitution and complementary effects might operate on the same consumer simultaneously. We cannot identify these two competing effects separately, but with our unique data, we can study the net effect of fixed-line Internet speed on mobile Internet adoption Extent of Usage To investigate the extent of use of the mobile Internet, conditional on adoption, we focus on two types of mobile phone activities. That is, consumers use mobile phones to perform online activities that correspond with activities performed through the fixed-line Internet (e.g., check , search for information, social networking, read news, online banking, share photo/video) but also perform mobile-

7 6 specific offline activities without corresponding activities through fixed-line Internet connections (e.g., SMS, take photos, make videos). We primarily investigate the effects of local fixed-line and mobile Internet speeds on the use of mobile online activities, and then show that the effects do not appear for the use of mobile offline activities as a falsification test. For the usage of mobile online activities, similar to our predictions for probability of adoption, we anticipate that mobile Internet speed may be positively associated with usage intensity, because the returns that consumers achieve from consuming more data increase more with faster mobile Internet connection speeds. Conditional on adoption, consumers might view the fixed-line and mobile Internet as complements or substitutes for performing their mobile online activities. In a falsification analysis, we consider the use of mobile-specific offline activities, for which mobile Internet speed should not correlate with usage intensity, because mobile-specific offline activities do not require any Internet connection, and greater mobile Internet speed does not affect their returns. Similarly, the fixed-line and mobile Internet should be neither complementary nor substitutive for such activities, because they can only be performed on mobile phones and are not available through fixed-line networks. 3. Empirical Analysis 3.1. Model Adopting a new technology involves two decisions: adoption and the extent of use (Astebro 2004; Fichman and Kemere 1997). To investigate empirically how the local speed of mobile and fixed-line Internet services influences consumers probability of adoption and extent of use of the mobile Internet, we apply a logit model with two stages, in which consumers first choose whether to adopt a mobile data plan, then choose whether to use mobile activities more or less frequently, conditional on adoption. assumed to be The utility of adopting a mobile data plan in the first stage for individual i living in zip code z is

8 7 (1) = , where is the average fixed-line Internet download speed in z across all available carriers; is the average mobile Internet download speed available in z across all available carriers; 1 and measure the sensitivity of mobile Internet adoption to local fixed-line and mobile Internet speed. If mobile Internet speed has a positive impact on mobile Internet adoption, should be significantly positive. If the fixed-line and mobile Internet are substitutes, the effect of the fixed-line speed on mobile Internet adoption should be negative, <0, whereas if they are complements, the effect of the fixed-line speed on mobile Internet adoption should be positive, >0. Many other factors may influence the utility of adopting a mobile data plan too. are individuallevel control variables, such as age, gender, income, area, the number of people in the household, and the length of the mobile service usage; 2 are the zip code level demographic control variables, including median income, population, and percentage of the population with a bachelor s degree; 3 is the number of active fixed-line service providers in zip code z; are a list of dummy variables indicating whether a mobile carrier j is available in zip code z; We also include a series of state dummies indicating whether individual i lives in state k,, and an inclusive value,, calculated from the second stage of the logit model. 1 The results remain consistent if we use the average of download and upload speeds for fixed-line and mobile Internet. 2 Age, gender, education, income and area are coded as categorical variables in our data. For example, the categories for the age group variable are 2=18 24, 3=25 34, 4=35 44, 5=45 54, 6=55 64, and 7=65 and over; the categories for gender are 0=Female, 1=male; the categories for the income-level variable are 1=<$25,000, 2=$25,000 $50,000, 3=$50,000 $75,000, 4= =$75,000 $100,000, and 5=greater than $100,000; the categories for education level are 1=grammar school or less, 2=some high school, 3=high school completed, 4=some college, 5=associate degree, 6=bachelor s degree, and 7=postgraduate degree; the categories for area are 0=Rural, 1=Urban. 3 We obtain the zip code level demographic data from the 2010 U.S. Census.

9 8 However, an important omitted variable in our model is price. After careful and extensive searching, we were not able to find price data for either fixed-line Internet or mobile Internet services in our sample period. Different providers react differently to similar economic conditions and determinants, such as regulatory incentives and demographic compositions of users, and thus offer widely varying Internet services plans with diverse prices and speeds (Greenstein and Mazzeo 2006). Although the price of a service plan offered by a carrier does not vary across markets (German 2013), the available service plans and speeds vary geographically. Unfortunately, we are unable to observe the specific service plans offered in local geographic markets in our sample. Because the service plans offered reflect the active carriers in a market, we include a series of mobile carrier dummy variables, as controls in our empirical model to indicate whether a carrier j is available in zip code z. Similarly, because the service plans offered depend on the level of competition (Dai 2008; Greenstein and Mazzeo 2006), we control for the number of active fixed-line Internet providers in a local market with. 4 Even with these controls, we might not account for all aspects of the service quality offered by carriers in a local geographic market. Thus, the variables and should be viewed in part as proxies for the overall level of service provided across the two platforms. Another set of omitted variables are local telecommunications regulatory rules, which could influence mobile service decisions, because of different firms unique reactions to similar regulatory incentives (Greenstein and Mazzeo 2006). Because regulatory rules are normally applied at the state level, we use the state dummies in the model to control for any specific local telecommunications regulatory rules. As we note in detail subsequently, the state dummies also control for potential differences in the way we collected service data across states. 4 We do not control for the number of active mobile carriers, because the number of active mobile carriers does not vary as much across areas as the number of active fixed-line providers. Unlike for the mobile carriers, we do not include dummy variables to indicate which fixed-line Internet provider is active in a local market, because there are 14 possible fixed-line providers in our data, and our model cannot be identified if we include the large number of dummy variables that would be necessary to identify these parameters.

10 9 In the second stage, conditional on mobile Internet adoption, consumers choose whether to use mobile activities more or less frequently. Specifically, the utility of choosing to perform mobile activities more frequently, conditional on mobile data plan adoption by individual i at zip code z, is assumed to be (2) = , where and measure the sensitivity of mobile usage to local fixed-line and mobile Internet If mobile Internet speed has a positive impact on usage of mobile activities, should be significantly positive. If the fixed-line and mobile Internet are substitutes, the effect of the fixed-line speed on usage of mobile activities should be negative, <0; if the fixed-line and mobile Internet are complements, the effect of the fixed-line speed on usage of mobile activities should be positive, >0. We include the same control variables from the Equation (1). We conducted separate analyses of the uses of mobile online activities and mobile-specific offline activities using the model represented by Equation (2). We first examine the effects of fixed-line and mobile Internet speeds on the use of mobile online activities, and then show in a falsification test that there are no effects of fixed-line and mobile Internet speeds on the use of mobile-specific offline activities Data To estimate our empirical models, we used data from the second quarter of 2011, obtained from comscore MobiLens. The MobiLens data are constructed from detailed, quarterly surveys of the mobile activities of a nationally representative sample of U.S. mobile subscribers, who have been recruited from a large, demographically diverse population through a variety of recruitment methods. Each quarterly sample includes responses from more than 20,000 individual mobile subscribers. The survey asks mobile phone users about their consumption behaviors across a wide variety of mobile activities, including

11 10 whether they adopted a mobile data plan and usage frequencies for different mobile online and mobilespecific offline activities. 5 We combined the MobiLens data with local Internet coverage data (fixed-line and mobile Internet) obtained from the National Broadband Map (NBM). The National Broadband Map data (NBM) are generated and managed by the U.S. National Telecommunications and Information Administration, with assistance from the Federal Communications Commission (FCC). These data were first released in June NBM data have been used to compare broadband availability across geographic areas and demographic groups, which can inform policies to support private sector investments in deploying broadband (National Telecommunications and Information Administration, 2011). The NBM data are collected by individual, separate state agencies, and we can use the state fixed effects in our model to control for any differences in data collection processes. The finest geographic level in the comscore MobiLens data is a consumer zip code; in contrast, the NBM data are reported at the census block level. We matched zip codes with census blocks using the point-in-polygon method by determining if the centroid of a block fell within a polygon of a zip code (Gordon-Larsen et al. 2006; Henry and Boscoe 2008). We also conducted robustness checks with other criteria, including whether the centroid of a block was within a certain radius range (20 miles, 30 miles, 50 miles) of the centroid of a zip code. The results remained consistent. Summary statistics are shown in Table 1. (Insert Table 1 Here) To proxy for the extent of use of mobile online activities, conditional on mobile Internet we applied cluster analysis and classified consumers into two clusters, according to whether they performed mobile online activities frequently. We would like to examine the usage intensity across all mobile online activities, rather than any particular category of activity. To do this, we aggregated usage 5 Usage frequencies are categorical variables in our data: never used this service, used it before but not in the previous month, used it once to three times in the previous month, used it at least once each week in the previous month, or used it almost every day in the previous month.

12 11 frequency variables for a list of different categories of mobile online activities in our data (e.g., , search for information, social networking, read news, online banking, share photo/video) using a k- medians clustering algorithm. 6 Specifically, on the basis of the usage frequencies of n categories of mobile online activities, we partitioned all consumers into two clusters, such that each consumer belonged to the cluster with the nearest median vector. 7 Thus, one cluster of consumers frequently used all categories of mobile online activities, and the other used all categories of mobile online activities much less frequently. Similarly, to classify consumers on the basis of whether they perform mobile-specific offline activities frequently, we used usage frequency variables for different mobile-specific offline activities (e.g., SMS, take photos, make videos) and conducted the cluster analysis according to the k-medians algorithm. The cluster of consumers who performed mobile online activities more frequently did not correlate perfectly with the cluster of consumers who performed mobile-specific offline activities more frequently, so a consumer might engage in neither mobile online activities nor mobile-specific offline activities Identification Strategy A relationship between Internet speed and mobile Internet adoption and usage does not, in itself, mean that speed causes adoption and usage. For example, service providers might choose to improve service because mobile Internet adoption and usage is growing faster in a particular location. More importantly, there may be omitted variables that could be correlated with both Internet speed and mobile service adoption and usage, such as advertising expenditures by service providers. The costs of delivering Internet service differ in ways that do not correlate with consumers adoption and usage behavior (Kolko 2012). We use this fact to identify the effect of mobile and fixed-line 6 The results remain consistent with other clustering algorithms, such as k-means clustering or hierarchical clustering algorithms with different linkage criteria. 7 The Calinski Harabasz pseudo-f statistics affirm that this two-group clustering (more vs. less frequent) is more effective than using more than two groups, which indicates that consumers usages correlate highly across different categories of activities.

13 12 Internet quality on mobile Internet adoption and usage, We instrumented local mobile Internet speed by using historical coverage information in adjacent areas, namely, the number of mobile carriers that offered the service six years ago in areas adjacent to the focal zip code (Dai 2008). 8 An adjacent area is a zip code within the same Census city. We also conducted robustness checks with different definitions of adjacent areas such as zip codes in the same county or within a certain radius range (50 miles, 30 miles) and found consistent results. Because of the economies of scale or scope associated with providing mobile service in a focal area (Boer and Evans 1996), there is a high spatial correlation in installation and maintenance costs for mobile services across adjacent locations. Moreover, preexisting service by a carrier influences its current service plan, because of the sunk costs of its investment (Czernich et al. 2011). Therefore, the number of carriers offering service in adjacent areas six years prior should be closely associated with the quality of mobile service offered in the focal area, but the available mobile service offered six years ago in adjacent areas should not correlate with consumers adoption or usage decisions in our sample period. Next, we instrumented for fixed-line Internet speed with a proxy of the engineering cost of wire service in the local exchange area, calculated according to the FCC s Hybrid Cost Proxy Model (HCPM) as of January The HCPM is an economic engineering model that calculates the cost of providing local wire service using efficient technology, given an area s geographic terrain and subscriber density (Prieger 2003). Proxy costs were not available for approximately one-third of the wire centers (mostly smaller carriers), so in these cases, we used the cost for the nearest wire center available (Prieger 2003). Wire center boundaries were matched to zip code areas. The costs of providing local wire service correlated strongly with the quality of the provider s fixed-line Internet services but were unlikely to correlate with consumers mobile Internet adoption or usage decisions. 8 Mian Dai from Drexel University for provided us with historical data on mobile carrier service, collected in December 2005 from letstalk.com, a leading online wireless service retailer. Our choice of using a six year lag for our instrument is because of the time period when these data were collected and our sample year (2011). 9 James E. Prieger from Pepperdine University provided us these proxy cost data for wire services in the local exchange area.

14 13 We used a two-step control function approach, with the two instrumental variables, to address endogeneity concerns, namely, and might correlate with the error term in the model (Imbens and Wooldridge 2007; Wooldridge 2014). A two-step control function approach aims to decompose the error term into a part that can be explained by a linear function of endogenous variables and the remaining residual (Petrin and Train 2010). In the first step, we ran two separate linear regressions for the endogenous variables, mobile Internet speed and fixed-line speed, using two instruments, the number of mobile carriers offering service in adjacent areas six years ago and proxy costs of providing local wire service, as well as the entire set of exogenous variables in each equation. Thus we obtained two sets of estimated residuals, and. Next, we estimated the two-stage logit model simultaneously using standard maximum likelihood techniques, with all explanatory variables in the model plus the estimated residuals from the regressions in the previous step as extra explanatory variables. Specifically, the adoption utility model for individual i in zip code z becomes: = The utility model for usage is similar to the adoption utility model but excludes the inclusive value. To further verify that our findings were not driven by potential endogeneity concerns, we examined how the local fixed-line and mobile Internet speeds influenced the usage of mobile-specific offline activities, conditional on mobile Internet adoption. We did this as a falsification test: mobile Internet speed should have little impact on the usage of mobile-specific offline activities, because their consumption does not require a mobile Internet connection. Similarly, fixed-line Internet speed should not affect a consumer s use of mobile-specific offline activities because mobile-specific offline activities cannot easily be performed on the fixed-line Internet.

15 14 A potential concern associated with our Internet speed data is that the fixed-line Internet and mobile Internet speeds might vary consistently across different locations. If they are perfectly correlated, we can identify only the net effects, not the distinct effects of fixed-line and mobile Internet speed. However, the Spearman rank correlation between the fixed-line Internet and mobile Internet speeds across different U.S. locations in our data was 0.32, which indicated that these speeds did not correlate strongly. 4. Results 4.1. Main Results We presented the estimated coefficients from the models as represented by Equation (1) and (2) in Table 2, Panel A both without and with control functions. The coefficients of the number of carriers offering service in adjacent areas six years prior and the proxy costs to provide wire line service were significant (at 1%) when we instrumented them on the fixed-line Internet speed and mobile Internet speed, suggesting they are strong instruments. 10 A comparison of the models with and without the use of control functions revealed that the parameters of interest the positive coefficient of Speed and the negative coefficient of Fixed-line Speed become less positive when control functions are used, consistent with the anticipated omitted variable bias. Locations with high quality Internet service likely differ in unobserved ways that correlate positively with Internet adoption and usage, which in turn should cause a positive coefficient bias. Because the estimated coefficients reflect a non-linear logit model, in Panel B of Table 2 we reported the corresponding average marginal effect of a one standard deviation change in the variable of interest on the probability of choice across all observations in the sample, computing the average marginal effects across all covariate values in the distribution. Column 1 in Panel B showed the implied average marginal effects for the mobile data plan adoption decision from the specification in Column 3 of Panel A. The positive marginal effect of Speed suggested that better local mobile Internet coverage increased the likelihood of adopting mobile data plans; the negative 10 The estimation results of the first step of the control function approach that is used to compute the residuals for the endogenous variables are in Table A1 in the Appendix.

16 15 marginal effect of Fixed-line Speed suggested that if the local fixed-line connection was insufficient, consumers got online through their mobile phones as a substitute. Specifically, a one standard deviation (0.59 mbps) increase in mobile Internet speed leaded to a percentage point increase in the probability of choosing a mobile data plan, on average. The overall average probability of adopting a mobile data plan was 0.35, so this increase represented a 29.77% increase. In contrast, a one standard deviation (32.87 mbps) increase in the fixed-line Internet speed decreased the probability of choosing a mobile data plan by 8.01 percentage points, equivalent to a 22.86% drop. In Column 2 of Table 2, Panel B, the significantly positive marginal effect of Speed and significantly negative marginal effect of Fixed-line Speed also affirmed that mobile Internet speed was positively associated with usage intensity of mobile online activities, but fixed-line Internet speed was negatively associated with this intensity, conditional on adoption. If mobile Internet speed increased by one standard deviation (0.59 mbps), consumers were more likely, by 9.87 percentage points, to conduct their mobile online activities more frequently. The overall probability of performing mobile online activities more frequently in our data was 0.28, so this increase was equivalent to 35.25%. If the fixedline Internet speed increased by one standard deviation (32.87 mbps), consumers were 8.22 percentage points (29.36%) less likely to conduct frequent mobile online activities though. (Insert Table 2 Here) 4.2. Robustness Checks In our falsification analysis, we applied the previously described two-stage logit model. Columns 5 and 6 in Panel A of Table 2 included the estimation results for the impacts of fixed-line and mobile Internet speeds on mobile data plan adoption and mobile-specific offline usage, conditional on data plan adoption. The average marginal effects for Speed and Fixed-line Speed are reported in Columns 3 and 4 in Panel B of Table 2. As we expected, the insignificant coefficient of Speed in Column 6 of Panel A suggested that mobile Internet speed did not significantly affect usage of mobile-specific offline activities. The insignificant coefficient of Fixed-line Speed in Column 6 of Panel A also confirmed that

17 16 fixed-line Internet speed did not affect usage of mobile-specific offline activities. The fixed-line Internet platform was not a good substitute for mobile-specific offline activities. These results are consistent with the predictions of our theoretical framework but not with an explanation based on the effects of the omitted variables correlated with speed. To further probe the effects of omitted variable bias on our results, we added more zip code level demographic control variables (percentage of men in a zip code, percentage of people younger than 45 years in a zip code, percentage of homeownership in a zip code). As we showed in Table 3, adding these controls did not change the estimates much; the impact of unobservable variables would have to great, relative to the impact of the observable variables, if omitted variables were to influence the results (Altonji, Elder, and Taber 2005). (Insert Table 3 Here) Although we cannot completely eliminate potential omitted variable bias, it is worth noting that many sources of potential omitted variables would tend to bias the coefficients of the fixed-line Internet speed positively. That is, places where fixed-line Internet speed is high likely are places where mobile Internet adoption and usage are high, for other unobserved reasons. For example, if more intense regional Internet adoption leads to both local Internet service improvements and more mobile Internet adoption and usage (Belo and Ferreira 2012), the fixed-line Internet speed would be positively associated with mobile Internet adoption and usage. Therefore, the negative impacts of fixed-line Internet speed on mobile Internet adoption and usage even might be underestimated due to the potential omitted variable bias Interaction Effects Local and Consumer Characteristics with Substitution We have reported that the speed of the fixed-line Internet has a negative impact on mobile Internet adoption and usage. We next identify consumers who will show stronger or weaker substitution between the mobile Internet and the fixed-line Internet, if any.

18 17 On average, the fixed-line Internet offers higher speeds than the mobile Internet (Federal Communications Commission 2013). In the NBM data for example, the average fixed-line Internet speed was mbps, and the average mobile Internet speed was 1.87 mbps in the second quarter of But in areas with relatively low fixed-line Internet speeds, these speeds are more comparable, which may make the fixed-line and mobile Internet closer substitutes for consumers when they perform online activities. The returns on adopting and using the mobile Internet for consumers in areas with relatively lower fixed-line Internet speeds increase, and they are more likely to rely on the mobile Internet to perform online activities, because the connection speed is similar to that offered by the poor fixed-line Internet. In short, the fixed-line and mobile Internet are closer substitutes when fixed-line speeds are low. We next consider the effects of a consumer s age on the degree of substitution. Younger users exhibited a stronger substitution effect between the mobile and fixed-line Internet. Although the mobile Internet generally is slower than the fixed-line Internet, younger consumers might be less sensitive to this slower speed, because their opportunity cost for time is lower than that for older people (Goldfarb and Prince 2008; Krantz-Kent and Stewart 2007; Posner 1995). Moreover, younger consumers likely use the features of a mobile phone more effectively and can compensate for the smaller screen when conducting online activities. Therefore, younger consumers may view the mobile Internet as a closer substitute for the fixed-line Internet than relatively older consumers. To capture the relative quality of the local fixed-line Internet, we created a dummy variable for each zip code, LowFixedSpeed z, which is equal to 1 if the fixed-line Internet connection speed was lower than the 25th percentile of the fixed-line Internet speed across all zip codes in the sample (7.7 mbps), and 0 otherwise. To capture consumers ages, we created another dummy variable for each individual, Young i, equal to 1 if the consumer s age category was either or 25 34, as represents the 25th percentile of the age categories across the entire sample, and 0 otherwise. Because some of the key moderators may correlate, we focused on the full utility models for both adoption and usage, including

19 18 both interaction variables. Specifically, the utility of individual i in zip code z of adopting a mobile data plan is: (3) = , and the utility model for usage intensity was similar but did not feature the inclusive value. The estimated coefficients of the model represented by Equation (3) for mobile Internet adoption and usage appear in Table 4, Panel A. We used the number of carriers offering service in adjacent areas six years prior, proxy costs to provide wire line service, and their interactions with Young i and LowFixedSpeed z as instruments for Fixed-line Speed, Speed, Fixed-line Speed Young, Speed Young, Fixed-line Speed Low Fixed-line Speed Area, and Speed Low Fixed-line Speed Area in the control function. Panel B of Table 4 contains the average marginal effects of a one standard deviation change in the variable of interest on the probability of choice across all observations by computing the average marginal effects across all covariate values in the distribution. The marginal effects revealed that the effects of the fixed-line Internet speed on mobile Internet adoption and usage were moderated by variation in both fixed-line Internet speed and consumer age. (Insert Table 4 Here) The coefficients of Fixed-line Speed Young were negative and statistically significant in both Column 1 and 2 of Table 4, Panel A. These results suggest that younger consumers will display stronger substitution between the mobile and fixed-line Internet. Specifically, as shown in the average marginal effects of Fixed-line Speed Young in Table 4, Panel B, given the same one standard deviation decrease (32.87 mbps) in the fixed-line Internet speed, the increase in the probability that younger consumers choose a mobile data plan was percentage points (28.74%) greater than the probability that other consumers did so. The increase in the probability of performing mobile online activities more frequently,

20 19 conditional on adoption, was percentage points (36.21%) higher for younger, compared with older, consumers. The coefficients of Fixed-line Speed Low Fixed-line Speed Area also were negative and statistically significant in both Columns 1 and 2 of Table 4, Panel A, suggesting that that in areas with lower fixed-line Internet speeds, consumers were more likely to adopt the mobile Internet and perform mobile online activities, substituting for the slow fixed-line Internet connection. To put these results in perspective, in Table 4, Panel B, we find that, according to the average marginal effects of Fixed-line Speed Low Fixed-line Speed Area, the same one standard deviation increase (32.87 mbps) in fixed-line Internet speed decreased the probability of choosing a mobile data plan among consumers in areas with higher average fixed-line Internet speed by 8.62 percentage points (24.62%) less than it did among consumers in areas with lower fixed-line Internet speeds. The decrease in the probability of performing mobile online activities more frequently, conditional on adoption, was 9.13 percentage points (32.61%) less for consumers in areas with higher, rather than lower, fixed-line Internet speeds. In contrast, the estimated coefficients of Speed Low Fixed-line Speed Area in both Columns 1 and 2 were insignificant in Table 4, Panel A and B. The impact of mobile Internet speed on mobile internet adoption and usage did not depend on different local fixed-line Internet speeds, as we expected. Furthermore, the estimated coefficients of Speed Young were insignificantly negative, which offers an interesting result. Although young consumers appear more sensitive than older consumer to changes in fixed-line speed, mobile Internet speed had equivalently positive impacts on mobile Internet adoption and usage by consumers of all ages. One possible explanation for this result is that younger consumers preferences for mobile Internet access make their behavior less sensitive to changes in mobile service quality. In support of this view, Table 4, Panel A, indicates that younger consumers are more likely to adopt and use all aspects of the mobile Internet, all else being equal. However, more investigation into this question is needed. Finally, we found no significant interaction effects for mobilespecific offline usage, as we expected.

21 Robustness Checks for Interaction Effects To determine age and area effects separately, instead of using dummy variables in a full model, we applied the two-stage logit model to the subpopulations of our sample. Columns 1 and 2 of Table 5, Panel A show the estimation results for younger consumers (<35), and the estimation results for consumers older than 35 are listed in Column 3 and 4. Column 5 and 6 include the estimation results for areas with lower fixed-line Internet speeds (<7.7 mbps), and Column 7 and 8 show the results for areas with higher fixed-line Internet speeds. The average marginal effects of the variables of interest are in Table 5, Panel B. Results for younger and older consumers are included in Columns 1 to 4, and those for areas with lower and higher fixed-line Internet speeds in Column 5 to 8. The coefficients of Fixed-line Speed for younger consumers were more negative than those for older consumers. Similarly, the coefficients of Fixed-line Speed for areas with lower fixed-line Internet speeds were more negative than those in areas with higher fixed-line Internet speeds. That is, younger consumers and consumers in areas with lower fixed-line Internet speeds exhibited stronger substitution effects between the fixed-line and mobile Internet, consistent with the results in Table 4. (Insert Table 5 Here) We examined the robustness of our results to different cutoff points to define the dummy variables for younger consumers and lower fixed-line Internet speed areas. To do this, we conducted another analysis based on a spline model for each variable, namely, with interactions for all the fixed-line Internet speed percentile dummies (10%, 25%, 50%, 75%, 90%) and interactions for all age categories (18 24, 25 34, 35 44, 45 54, 55 64, 65+ years). The estimated coefficients for the interactions between fixed-line speed and the percentile dummies of fixed-line Internet speed were significantly negative (i.e., stronger substitution) at 10% and 25% but insignificant for other percentiles. The estimated coefficients for the interactions between fixed-line speed and the age categories were significantly negative (i.e., stronger substitution) for and years but insignificant for the other age categories. 11 Thus, 11 These detailed results are available on request.

22 21 consumers younger than 35 years and those living in areas with a fixed-line Internet speed lower than the 25th percentile exhibited stronger substitution between the two Internet platforms. 5. Discussion and Conclusion 5.1. Summary With this study, we have sought to investigate how the quality of a possible substitute influences the adoption and usage of the mobile Internet. Prior research on mobile Internet consumer behavior has focused on usage patterns, conditional on adoption of a mobile Internet service (e.g., Ghose and Han 2011; Ghose, Goldfarb, and Han 2012; Xu et al. 2014a; Xu et al. 2014b), but we know less about consumers decisions across different Internet platforms. Considering the potential for multichannel consumption behavior, we tested for complementarity or substitution across different Internet platforms, then established conditions in which complementarity or substitution was stronger. In contrast with arguments that the capabilities of fixed-line and mobile Internet platforms are sufficiently different in many ways, which might lead to a potential expectation of complementarity between the two platforms, we found, using large-scale survey data and public Internet speed data, robust empirical evidence of a negative impact of local fixed-line Internet speed on mobile Internet adoption and usage. Thus, we offer evidence of substitution between the fixed-line and mobile Internet. This substitution effect is stronger among younger consumers, likely because these consumers are less sensitive to slower mobile Internet speeds and more accepting of the features of mobile devices. Because fixed-line and mobile Internet speeds are comparable in areas with relatively low fixed-line Internet speeds, we also show that consumers living in such areas exhibit stronger substitution between the fixed-line and mobile Internet. Better local mobile Internet coverage also could lead to a higher likelihood of adopting and using mobile Internet services Managerial and Policy Implications

23 22 The question of whether fixed-line and mobile Internet are substitutes has important implications for public policy. Telecommunication policy makers worry that market structures for mobile service provision are very concentrated. Most large mobile service providers face minimal competition from just a few other large mobile carriers. In introducing its National Broadband Plan to improve Internet access, the FCC (2010) showed that 98% of the U.S. population lived in census tracts with at least one 3G mobile provider, of which 77% lived in tracts with three providers, 12% had two, and 9% had one. Similarly, a lack of competition characterizes the fixed-line Internet market. According to the National Broadband Plan, most households can obtain fixed-line service from only one of two large firms in their area: 78% can choose between two providers, 13% have access to one, and 5% have none. Only 4% of the U.S. population lives in a location with three or more suppliers (FCC 2010). However, we find robust evidence that the mobile Internet serves as a substitute for the fixed-line Internet, and consumers switch platforms if they do not receive sufficiently fast data delivery. The resulting competition between mobile carriers and fixed-line Internet providers should alleviate some concerns in both the mobile and the Internet markets. Our findings also have important implications for mobile service providers. Carrier infrastructure investments are among the largest cost outlays for mobile providers; investments in mobile infrastructure, though costly, improve service quality, encourage consumer adoption of mobile services, and enhance post-adoption usage. Therefore, if a mobile service provider invests more to achieve better connection speed, more consumers likely adopt its Internet service. Our finding that fixed-line and mobile Internet are stronger substitutes in areas where the fixed-line Internet speed is low indicates that it will be more valuable for mobile carriers to provide higher mobile Internet speeds in these areas, to attract more consumers. These insights should inform telecommunication carriers strategic mobile infrastructure investments. With the emergence of a potentially disruptive platform, marketers must monitor changes in consumer behavior to understand their implications for marketing strategy (Ahonen 2011). For example,

24 23 advertisers may need to adjust their cross-platform advertising strategies to reach their target audience more effectively. We find that consumers go online through the mobile Internet if the fixed-line Internet is not satisfactory. This finding suggests that it will be harder for advertisers to reach their target audience, which likely is more dispersed across platforms. Advertisements placed across channels likely produce a single impression on a broader audience, rather than repeated impressions on the same 5.3. Limitations and Further Research Several important data and econometric limitations arise in this study. First, because we do not observe individual behaviors on the fixed-line Internet, we investigate how fixed-line Internet speed influences mobile Internet adoption and usage; we cannot examine how mobile Internet speed influences consumer behavior on the fixed-line Internet. Second, we know whether a consumer has adopted a mobile data plan, but we cannot observe the particular type of mobile data plan each consumer adopts. Although consumers choices of data plans likely influence the extent of their usage, we expect that consumers who frequently engage in mobile online activities adopt mobile data plans with higher bandwidth caps. Third, we observe consumers usage frequencies across different mobile online activities but do not know whether they use their mobile data plans or Wi-Fi to perform online activities. However, if consumers use their mobile phones to get online through available Wi-Fi, rather than a mobile data plan, fixed-line Internet speeds may be positively associated with usage of mobile online activities, because the speed of the Wi-Fi connection at home or at work depends on the speed of the associated fixed-line network. The possibility that consumers perform mobile online activities through Wi-Fi could bias the substitution effect between the fixed-line and mobile Internet toward greater complementarity. Therefore, the substitution we find might be underestimated. Our substitution finding also offers several directions for research. Considering prior findings of both complementarity and substitution, in different contexts, we call for additional studies that address usage of the fixed-line and mobile Internet to perform different online activities. If researchers could access data about Internet consumption behavior for both mobile and fixed-line settings, they could

25 24 identify when substitution or complementarity was likely to prevail for the usage of different types of online activities. For example, consumers might use both Internet platforms as substitutes to check and sent , because s generally are not data intensive, and either platform is sufficient. However, they might read online news on both platforms as complements, in that they prefer to check timely headlines with a mobile app on their smaller mobile screen but then further explore interesting news stories in-depth on their personal computers using a fixed-line Internet. Researchers also might extend the current study by investigating the mechanism or source of substitution. We note some similarities and dissimilarities between the two Internet platforms, but the MobiLens data cannot reveal the mechanisms responsible for the substitution we find. If researchers could access Internet consumption data together with data about the specific contents consumed on each platform, they could provide more detailed insights into consumer behavior. The different consumption patterns on the two Internet platforms also might evolve differently over time, as consumers become more familiar with using mobile devices to conduct online activities. Findings along these lines could help Internet content providers make more informed content management decisions across different Internet platforms.

26 25 References Ahonen, Tomi T. (2011) All the Numbers, All the Facts on the Trillion-Dollar Industry. Why Is Google Saying: Put Your Best People on? (accessed November 26, 2012), [available at trillion-dollar-industry-why-is-google-sayng-put-your-b.html]. Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski, Richard Lutz, Alan Sawyer, and Stacy Wood (1997) Interactive Home Shopping: Consumer, Retailer, and Manufacturer Incentives to Participate in Electronic Markets. Journal of Marketing. 61(July): Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber (2005) Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools. Journal of Political Economy. 113(1): Astebro, Thomas (2004) Sunk Costs and the Depth and Probability of Technology Adoption. Journal of Industrial Economics. 52(3): AT&T (2011) AT&T Ends Bid To Add Network Capacity through T- USA Purchase. (accessed September 26, 2014), [available at Avery, Jill, Thomas J. Steenburgh, John Deighton and Mary Caravella (2012) Adding Bricks to Clicks: Predicting the Patterns of Cross-Channel Elasticities over Time. Journal of Marketing. 76(May): Bang, Youngsok, Dong-Joo Lee, Kunsoo Han, Minha Hwang and Jae-Hyeon Ahn (2013) Channel Capabilities, Product Characteristics, and the Impacts of Channel Introduction. Journal of Management Information Systems. 30(2): Belo, Rodrigo and Pedro Ferreira (2012) Using Randomization Methods to Identify Social Influence in Networks. Working paper, Carnegie Mellon University. Boer, David Boles De and Lewis Evans (1996) The Economic Efficiency of Telecommunications in a Deregulated Market: The Case of New Zealand. Economic Record. 72(216): Bresnahan, Timothy F. and Manuel Trajtenberg (1995) General Purpose Technologies Engines of Growth. Journal of Econometrics. 65(1): comscore Inc. (2012) 2012 Future in Focus. Review Report Series, comscore Whitepaper. Czernich, Nina, Oliver Falck, Tobias Kretschmer and Ludger Woessmann (2011) Broadband

27 26 Infrastructure and Economic Growth. Economic Journal. 121(552): Dai, Mian (2008) Strategic Interactions and Market Structure in the U.S. Telecommunications Industry. Working Paper, Drexel University. David, Paul A. (1969) A Contribution to the Theory of Diffusion. Stanford Center for Research in Economic Growth, Stanford University. Federal Communications Commission (2010) Internet Access Services: Status as of December 31, (accessed April 23, 2014), [available at (2013) Connecting America: The National Broadband Plan. (accessed Jun 23, 2014), [available at Fichman, Robert G. and Chris F. Kemerer (1997) The Assimilation of Software Process Innovations: An Organizational Learning Perspective. Management Science. 43(10): Forman, Chris, Anindya Ghose, and Avi Goldfarb (2009) Competition Between Local and Electronic Markets: How the Benefit of Buying Online Depends on Where You Live. Management Science. 55(1): Gentzkow, Matthew (2007) Valuing New Goods in a Model with Complementarity: Online Newspapers. American Economic Review. 97(3): German, Kent (2013) Quick Guide to Cell Phone Carriers. (accessed June 23, 2014), [available at Ghose, Anindya, Avi Goldfarb, and Sang Pil Han (2012) How is the Internet Different? Search Costs and Local Activities. Information Systems Research. 24(3): and Sang Pil Han (2011) An Empirical Analysis of User Content Generation and Usage Behavior on the Internet. Management Science. 57(9): ,, and Kaiquan Xu (2013) Battle of the Channels: The Impact of Tablets on Digital Commerce. Working paper, New York University. Goldfarb, Avi and Jeff Prince Internet adoption and usage patterns are different: Implications for the digital divide. Information Economics and Policy 20(1): Gordon-Larsen, Penny, Melissa C. Nelson, Phil Page and Barry M. Popkin (2006) Inequality in the Built Environment Underlies Key Health Disparities in Physical Activity and Obesity. Pediatrics. 117(2): Greenstein, Shane and Michael Mazzeo (2006) The Role of Differentiation Strategy in Local

28 27 Telecommunication Entry and Market Evolution: Journal of Industrial Economics. 54(3): Henry, Kevin A. and Francis P. Boscoe (2008) Estimating the Accuracy of Geographical Imputation. International Journal of Health Geographics. 7(1):3. Imbens, Guido and Jeffrey Wooldridge (2007) Control Function and Related Methods. What s New in Econometrics, NBER Summer Institute. International Telecommunication Union (2013) The World in 2013: ICT Facts and Figures. ICT Data and Statistics Division. Karshenas, Massoud and Paul L. Stoneman (1993) Rank, Stock, Order, and Epidemic Effects in the Diffusion of New Process Technologies: An Empirical Model. RAND Journal of Economics, 24(4): Kolko, Jed (2012) Broadband and Local Growth. Journal of Urban Economics. 71(1): Krantz-Kent, Rachel and Jay Stewart (2007) How Do Older Americans Spend Their Time? Monthly Labor Review. 130(8):8 26. Limayem, Moez, Mohamed Khalifa, and Anissa Frini (2000) What Makes Consumers Buy from Internet? A Longitudinal Study of Online Shopping. Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions. 30(4): National Telecommunications and Information Administration (2011) NTIA Unveils National Broadband Map and New Broadband Adoption Survey Results. (accessed September 21, 2014), [available at Petrin, Amil and Kenneth Train (2010) A Control Function Approach to Endogeneity in Consumer Choice Models. Journal of Marketing Research. 47(1):3 13. Peukert, Christian and Tobias Kretschmer (2014) Video Killed the Radio Star? Online Music Videos and Digital Music Sales. Working Paper. Posner, Richard A (1995) Aging and Old Age. University of Chicago Press. Prieger, James E (2003) The Supply Side of the Digital Divide: Is There Equal Availability in the Broadband Internet Access Market? Economic Inquiry. 41(2): Wooldridge, Jeffrey (2014) Quasi-Maximum Likelihood Estimation and Testing for Nonlinear Models with Endogenous Explanatory Variables. Journal of Econometrics. 182(1):

29 28 Xu, Jiao, Chris Forman, Jun B. Kim and Koert Van Ittersum (2014) News Media Platforms: Complements or Substitutes? Evidence from Phone Usage. Journal of Marketing. 78(4): Xu, Lizhen, Jason A. Duan, Yu J. Hu, Yuan Cheng and Yan Zhu (2014) A Dynamic Structural Model for Heterogeneous Data Consumption and Targeted Promotion. Working paper, Georgia Institute of Technology. Yoo, Boonghee and Naveen Donthu (2001) Developing a Scale to Measure the Perceived Quality of an Internet Shopping Site. Quarterly Journal of Electronic Commerce. 2(1):31 45.

30 29 Table 1. Summary Statistics Variable Mean SD Min Max Adopt a mobile data plan Fixed-line speed (mbps/zip) speed (mbps/zip) Log median income ($/Zip) Log population (Zip) Percent bachelor's degree (Zip) Number of mobile carriers (Zip) Number of fixed-line Internet providers (Zip) Number of household members Length of mobile phone use (year) Number of observations 28,117 Notes: Summary statistics for Age, Gender, Education, Income, and Area are available on request.

31 30 VARIABLES Table 2. /Fixed-Line Internet Speed and Internet Adoption and Usage Main Two-Stage Model (No Control Function) A. Coefficient Estimates Main Two-Stage Model (With Control Function) Falsification Two-Stage Model (With Control Function) (1) (2) (3) (4) (5) (6) Data Plan Adoption Fixed-line Speed (0.0002) Speed ** (0.0423) Log Median Income (Zip) ** (0.0221) Log Population (Zip) ** (0.0231) Percent Bachelor's ** Degree (Zip) (0.0018) Number of Fixed-line Providers (Zip) * (0.0167) Online Usage * (0.0003) ** (0.0374) ** (0.0236) ** (0.0203) ** (0.0015) * (0.0092) Data Plan Adoption * (0.0003) *** (0.0358) *** (0.0071) *** (0.0098) *** (0.0007) * (0.0130) Online Usage ** (0.0004) ** (0.0274) ** (0.0125) * (0.0236) ** (0.0013) * (0.0068) Data Plan Adoption * (0.0003) *** (0.0171) *** (0.0083) *** (0.0089) *** (0.0006) * (0.0127) Demographic Controls Yes Yes Yes Yes Yes Yes Carrier Dummies Residual for Fixed-Line Speed Residual for Speed Yes Yes Yes Yes Yes Yes ** (0.0267) ** (0.0102) * (0.0199) ** (0.0123) ** (0.0228) ** (0.0110) - Specific Offline Usage (0.0002) (0.0326) ** (0.0110) (0.0247) (0.0014) (0.0128) * (0.0339) ** (0.0103) Inclusive Value (0.2836) (0.2677) (0.3074) Constants *** (0.9785) * (0.5142) *** (0.8313) * (0.4324) *** (0.8012) * (0.4475) Observations 28,117 28,117 28,117 28,117 28,117 28,117

32 31 B. Average Marginal Effects VARIABLES Main Two-Stage Model Falsification Two-Stage Model (1) (2) (3) (4) Data Plan Adoption Fixed-line Speed ** (0.0356) Speed ** (0.0423) Online Usage (Conditional on Adoption) ** (0.0427) ** (0.0394) Data Plan Adoption ** (0.0355) ** (0.0423) -Specific Offline Usage (Conditional on Adoption) (0.0321) (0.0747) *Significant at 10%; **Significant at 5%; ***Significant at 1%. Notes: The demographic controls were age, gender, income, area, number of people in the household, length of mobile service usage, and state dummies. Heteroskedasticity-robust standard errors appear in parentheses. In Panel A, the control functions with instrumental variables are in Column 3 6. A positive residual suggests that the mobile Internet possesses desirable attributes that are not controlled for by observable variables. Panel B presents the marginal effect of a one standard deviation change in each variable on the probability of the outcome, according to the average marginal effects across all covariate values in the distribution. The marginal effects for models with control functions are reported. Columns 1 4 in Panel B correspond to columns 3 6 in Panel A. The results remain robust to the marginal effects at the mean values of the covariates.

33 32 Table 3. /Fixed-Line Internet Speed and Internet Adoption and Usage: Additional Zip Code Level Controls Main Two-Stage Model Falsification Two-Stage Model VARIABLES (1) (2) (3) (4) Data Plan Adoption Fixed-line Speed * (0.0003) Speed *** (0.0158) Log Median Income (Zip) *** (0.0119) Log Population (Zip) *** (0.0104) Percent Bachelor's Degree (Zip) *** (0.0007) Percent Male (Zip) (0.0045) Percent Age < 45 (Zip) (0.0263) Percent Homeownership (Zip) * (0.0036) Number of Fixed-line Providers * (Zip) (0.0104) Online Usage ** (0.0004) ** (0.0211) ** (0.0226) * (0.0197) ** (0.0014) (0.0074) (0.0318) (0.0183) * (0.0058) Data Plan Adoption * (0.0003) *** (0.0159) *** (0.0203) *** (0.0076) *** (0.0007) (0.0046) (0.0248) * (0.0035) * (0.0111) Demographic Controls Yes Yes Yes Yes Carrier Dummies Yes Yes Yes Yes -Specific Offline Usage (0.0002) (0.0316) ** (0.0112) (0.0253) (0.0015) (0.0083) (0.0281) (0.0176) (0.0134) Residual for Fixed-Line Speed ** (0.0231) * (0.0165) ** (0.0247) * (0.0226) Residual for Speed ** (0.0100) ** (0.0115) ** (0.0118) ** (0.0092) Inclusive Value (0.2750) (0.3238) Constants *** (0.5729) * (0.2694) *** (0.6826) * (0.2347) Observations 28,117 28,117 28,117 28,117 *Significant at 10%; **Significant at 5%; ***Significant at 1%. Notes: The demographic controls were age, gender, income, area, number of people in the household, length of mobile service usage, and state dummies. Heteroskedasticity-robust standard errors appear in parentheses. We used control functions with instrumental variables. A positive residual suggests that the mobile Internet possesses desirable attributes that were not controlled for by observable variables.

34 33 Table 4. /Fixed-Line Internet Speed and Internet Adoption and Usage: Age and Fixed-Line Internet Speed Interaction Effects A. Coefficient Estimates Main Two-Stage Model Falsification Two-Stage Model VARIABLES (1) (2) (3) (4) Data Plan Adoption Fixed-line Speed * (0.0003) Online Usage * (0.0004) Data Plan Adoption * (0.0003) - Specific Offline Usage (0.0002) Speed *** (0.0180) *** (0.0110) *** (0.0169) (0.0347) Young ** (0.0128) Fixed-line Speed Young * (0.0003) Speed Young (0.0367) Low Fixed-line Speed Area * (0.0048) Fixed-line Speed Low Fixed-line * Speed Area (0.0003) Speed Low Fixed-line Speed Area (0.0095) Log Median Income (Zip) *** (0.0211) Log Population (Zip) *** (0.0165) Percent Bachelor's Degree (Zip) ** (0.0007) Number of Fixed-line Providers * (Zip) (0.0117) ** (0.0133) ** (0.0004) (0.0392) * (0.0051) * (0.0004) (0.0119) ** (0.0295) *** (0.0113) ** (0.0008) * (0.0063) ** (0.0121) * (0.0003) (0.0378) * (0.0056) * (0.0003) (0.0097) *** (0.0221) *** (0.0171) ** ( ) * (0.0120) Demographic Controls Yes Yes Yes Yes Carrier Dummies Yes Yes Yes Yes * (0.0099) (0.0049) (0.0274) (0.0272) (0.0196) (0.0317) ** (0.0326) (0.0289) (0.0043) (0.0142) Residual for Fixed-Line Speed * (0.0213) * (0.0204) * (0.0208) * (0.0296) Residual for Speed ** (0.0104) ** (0.0118) ** (0.0106) ** (0.0115) Residual for Interactions Yes Yes Yes Yes Inclusive Value (0.2835) (0.3410) Constants *** (0.3917) ** (0.6875) *** (0.3877) ** (0.3903) Observations 28,117 28,117 28,117 28,117

35 34 VARIABLES B. Average Marginal Effects Main Two-Stage Model Data Plan Adoption Fixed-line Speed ** (0.0195) Speed ** (0.0384) Fixed-line Speed Young ** (0.0473) Speed Young (0.0409) Fixed-line Speed Low Fixedline Speed Area Speed Low Fixed-line Speed Area ** (0.0419) (0.0386) (1) (2) Online Usage (Conditional on Adoption) ** (0.0232) ** (0.0337) ** (0.0496) (0.0441) ** (0.0465) (0.0374) *Significant at 10%; **Significant at 5%; ***Significant at 1%. Notes: The demographic controls were age, gender, income, area, number of people in the household, length of mobile service usage, and state dummies. Heteroskedasticity-robust standard errors appear in parentheses. We used control functions with instrumental variables. In Panel B, we present the marginal effect of a one standard deviation change in each variable on the probability of the outcome, computed according to the average marginal effects across all covariate values in the distribution. The effects are calculated for each individual, and then averaged across individuals in the sample. The results remain robust to marginal effects at the mean values of the covariates.

36 35 Table 5. /Fixed-Line Internet Speed and Internet Adoption and Usage: Age and Fixed-Line Internet Speed Split Samples VARIABLES Fixed-line Speed A. Coefficient Estimates Split Samples, Young Split Samples, Old Split Samples, Low Speed Split Samples, High Speed (1) (2) (3) (4) (5) (6) (7) (8) Data Plan Adoption * (0.0006) Online Usage * (0.0009) Data Plan Adoption (0.0003) Online Usage * (0.0003) Data Plan Adoption * (0.0005) Online Usage * (0.0008) Data Plan Adoption (0.0004) Online Usage * (0.0004) Speed *** (0.0124) *** (0.0103) *** (0.0275) ** (0.0374) *** (0.0281) *** (0.0193) *** (0.0227) ** (0.0315) Log Median Income (Zip) Log Population (Zip) Percent Bachelor's Degree (Zip) Number of Fixed-line Providers (Zip) Demographic Controls Carrier Dummies Residual for Fixed-Line Speed Residual for Speed Inclusive Value ** (0.0193) *** (0.0114) ** (0.0072) * (0.0135) ** (0.0216) *** (0.0136) ** (0.0080) * (0.0066) ** (0.0240) ** (0.0252) * (0.0020) * (0.0188) ** (0.0229) ** (0.0215) * (0.0020) * (0.0105) ** (0.0205) *** (0.0112) ** (0.0075) * (0.0133) ** (0.0228) *** (0.0153) ** (0.0068) * (0.0071) ** (0.0221) ** (0.0238) * (0.0020) * (0.0191) Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes * (0.0251) ** (0.0151) (0.2962) Constants *** (0.3182) * (0.0195) ** (0.0125) ** (0.2846) ** (0.0273) ** (0.0102) (0.3004) *** (0.8412) * (0.0172) ** (0.0132) * (0.3518) * (0.0213) ** (0.0144) (0.3216) *** (0.3918) * (0.0202) ** (0.0107) ** (0.2163) ** (0.0267) ** (0.0110) (0.3483) *** (0.8257) ** (0.0236) ** (0.0216) * (0.0019) * (0.0099) * (0.0184) ** (0.0129) * (0.3592) Observations 9,483 9,483 18,634 18,634 7,029 7,029 21,088 21,088

37 36 VARIABLES Fixed-line Speed B. Average Marginal Effects Split Samples, Young Split Samples, Old Split Samples, Low Speed Split Samples, High Speed (1) (2) (3) (4) (5) (6) (7) (8) Data Plan Adoption ** (0.0686) Online Usage ** (0.0739) Data Plan Adoption ** (0.0205) Online Usage ** (0.0232) Data Plan Adoption ** (0.0682) Online Usage ** (0.0704) Data Plan Adoption ** (0.0236) Online Usage ** (0.0279) Speed ** (0.0275) ** (0.0288) ** (0.0437) ** (0.0484) ** (0.0442) ** (0.0485) ** (0.0304) ** (0.0337) *Significant at 10%; **Significant at 5%; ***Significant at 1%. Notes: The demographic controls were age, gender, income, area, number of people in the household, length of mobile service usage, and state dummies. Heteroskedasticity-robust standard errors appear in parentheses. We used control functions with instrumental variables. Panel B presents the marginal effect of a one standard deviation change in each variable on the probability of the outcomes, computed as the average marginal effects across all covariate values in the distribution. The effects are calculated for each individual, and then averaged across individuals in the sample. The results remain robust to marginal effects at the mean values of the covariates

38 37 Appendix Table A1. Linear Regressions for the Endogenous Variables on Instruments VARIABLES Number of Carriers (0.1724) Proxy Cost *** (0.0130) Log Median Income (Zip) *** (0.1647) Log Population (Zip) *** (0.2635) Percent Bachelor's Degree (Zip) *** (0.0176) Number of Fixed-line Providers (Zip) * (0.0458) (1) (2) Fixed-Line Internet Speed Internet Speed Demographic Controls Yes Yes *** (0.0129) * (0.0034) *** (0.1048) *** (0.0127) ** (0.0019) (0.0583) Carrier Dummies Yes Yes Constants *** *** (1.2845) (0.0221) Observations 28,117 28,117 *Significant at 10%; **Significant at 5%; ***Significant at 1%. Notes: The demographic controls were age, gender, income, area, number of people in the household, length of mobile service usage, and state dummies. Heteroskedasticity-robust standard errors appear in parentheses.